On June 4, a screenshot from the cc4 Groups aggregator pricing list appeared on X (formerly Twitter): among the available models, a line was found claude-oceanus-v1-p. Price, context, latency – everything is in place. There is no official announcement from Anthropic. We break down what this means.
On social media, this model is being discussed under the name Claude Mythos — Oceanus is the internal identifier of a specific checkpoint currently undergoing red team testing.
Where the name came from and what the screenshot is
In the interface of the cc4 Groups aggregator, which collects models from different providers into a single API, a model appeared with the identifier claude-oceanus-v1-p.
The screenshot shows: Input 16 $/1M tokens, Output 80 $/1M, Cached 1,6 $. Context window 1M tokens, latency 9,16 s, speed 52 tokens per second. The status is partially red – the model is either not yet publicly active or in the process of rolling out.
The suffix -p by analogy with previous leaks means preview. This is exactly how the Claude Mythos Preview identifier looked before the official announcement.
Claude Oceanus and the model lineup: where it fits in
To understand the scale, context is needed. In March 2026, due to an error in the Anthropic CMS, drafts describing the Claude Mythos model were leaked to the public. The company confirmed the existence of the new development and called it a "qualitative leap" compared to everything that came before. In April, Mythos Preview was launched in closed mode via Project Glasswing – only for cybersecurity organizations. The model never reached public access: its capabilities in automating cyberattacks were too dangerous.
On May 28, Anthropic released Claude Opus 4.8 – an update to the flagship with improved coding and agent tasks. The announcement explicitly states: Opus 4.8 is inferior to Mythos, but "Mythos-level models are expected in the coming weeks".
Claude Oceanus appears immediately after this statement. A coincidence is unlikely.
What the $16/$80 price tag says
Let's compare with what is already known:
- Claude Opus 4.8 (public): $5/$25 per 1M tokens
- Claude Mythos Preview (closed, only Project Glasswing): $25/$125 per 1M tokens
- Claude Oceanus v1-p (leak): $16/$80 per 1M tokens
Oceanus is more expensive than the public Opus, but noticeably cheaper than the closed Mythos. It looks most like a public version of the Mythos class – retaining some capabilities, but without the most sensitive cyber tools that Anthropic is not ready to give to just anyone.
For comparison: GPT-5.6 is rumored to be preparing for release in the same time window. The price war between Anthropic and OpenAI is clearly trending downwards, and $16/$80 fits into this trend.
When to expect the announcement?
There are no exact dates. But there are two signals.
First: Anthropic publicly promised "Mythos-level models in the coming weeks" at the end of May.
Second: the model has already appeared in the aggregator's price list – which means the API is at least partially ready.
Both signals point to June 2026 as a realistic horizon.
Anthropic did not plan to make Mythos public at all. If Oceanus does come out for everyone – it is a reversal in strategy. Either the company managed to develop sufficiently strict guardrails, or it is a separate model with a different capability profile.
What this means for those who use Claude for work
If Oceanus is a public version of the Mythos class, the situation changes for several task categories.
Code audit and vulnerability scanning have so far been available only to Project Glasswing partners. If Oceanus is released to the open API, these tools will become available to any developer – without a lead and verification.
Long-horizon agentic tasks will get a more powerful foundation. Opus 4.8 already handles multi-step scenarios well, but Anthropic itself admits that the Mythos level is a different league.
The $16/$80 price tag does not look prohibitive. For professional use, this is a working range, especially considering caching.
So far, this is just a leak. Watching for the official announcement.
What happened after the leak
Events developed quickly. Shortly after the model appeared in the aggregator's price list, it turned out: one of the red team members resold access to claude-oceanus-v1-p via a Chinese proxy service. Anthropic suspended access for the entire group of testers and launched an internal investigation. On June 6, an identifier appeared in one of the AI trackers for a few minutes Claude-Mythos-5 – after which it was deleted. What is behind it is still unknown.
If you are just learning AI model terms – check out our glossary: it covers RAG, tokens, context window and other basic concepts.
